cerebro — Knowledge bases an AI keeps current for you
Most "chat with your documents" tools rediscover your files on every question. Nothing accumulates. A brain works differently: the AI reads each source once and maintains a living wiki that only grows.
Uploading a PDF to a chat
doesn't build anything
Most "chat with your documents" tools run on RAG: you store files, the AI retrieves relevant chunks on every question, and generates an answer. It works, but nothing accumulates. Ask something that spans five documents and the AI reassembles the pieces from scratch, every single time.
I wanted something different: concrete knowledge bases (business books, accounting, management) I could cite whenever I need to validate an idea or make a hard decision. Not a search engine over my PDFs. A memory that builds itself.
A wiki the AI
writes and maintains for you
A "brain" is a collection of interlinked Markdown pages built from your sources. The AI reads each book or PDF once, extracts what matters, and integrates it into the existing wiki: updates concept pages, cross-references, flags when a new source contradicts an older one. You just curate the sources and ask the questions.
The difference from RAG is that knowledge is compiled once and kept current, not re-derived on every query. Cross-references are already there. Contradictions are already flagged. Every question you answer makes the wiki a little richer.
Four operations.
All in plain language.
The whole system boils down to four actions you speak to your AI:
The most important rule lives in the query operation: if the brain doesn't cover something, it says so explicitly instead of inventing an answer and attributing it to your sources. That's the difference between a tool you can trust for decisions and one that hallucinates with confidence.
A normative document,
not just a skill
The core of this project isn't the Claude skill, it's BLUEPRINT.md: a document that specifies exactly how any AI should create, ingest, query, and lint a brain. The skill is just a thin wrapper around that document, which means the pattern survives any single vendor's format changes.
The repo also ships a complete example brain, built from two public-domain classics on money and business, so anyone can see the pattern working without installing anything.
Why it's on GitHub
and not a closed product
cerebro implements the "LLM Wiki" pattern described by Andrej Karpathy: the original idea, as a conceptual document. This repo turns it into an installable kit, with a normative blueprint, a ready-to-use skill, and multi-AI templates, something that was missing.
I didn't want another search engine over my PDFs. I wanted a brain that remembered what I'd already read and handed it back to me, cited, when I needed it.